Hybrid Approaches for Efficient Diffusion Models
Akie Kuro, Kaoru Suzu, Saburo Yuuta, Zoe Brian · HAL (Le Centre pour la Communication Scientifique Directe) · 2025
Diffusion models have emerged as a dominant paradigm in generative modeling, achieving state-of-the-art results in image synthesis, video generation, speech processing, and molecular design. Despite their impressive performance, the practical deployment of diffusion models is hindered by high computational costs, slow sampling times, and resource-intensive training. This survey provides a comprehensive overview of recent advancements aimed at improving the efficiency of diffusion models. We categorize efficiency improvements into four key areas: (1) accelerated sampling methods that reduce the number of function evaluations required for high-quality generation, (2) architectural optimizations that enhance computational efficiency while maintaining expressive power, (3) model compression and distillation techniques that enable lightweight deployments, and (4) alternative training and inference paradigms that reduce the overall complexity of diffusion models. We also explore the real-world impact of these optimizations across various application domains, including real-time image generation, speech synthesis, medical imaging, and drug discovery. Finally, we discuss open challenges and future research directions to further enhance the efficiency, scalability, and interpretability of diffusion models. By addressing these challenges, diffusion models can become more practical for deployment in resource-constrained environments and real-time applications.